Approaches to autonomous navigation in structured environments, such as road networks, can generally be divided into two categories. In the first category, the autonomous vehicle navigates like a human, with little pre-existing knowledge of the road features beyond simple maps and general rules of the road. The second category of approaches relies on extensive prior knowledge of the environment, for example, a dense set of global positioning system (GPS) measurements marking the lane boundaries and a map of all stop signs, pedestrian crossings and the like. The first category presents a challenging approach due to the extreme variability of real-world environments. The second category approach has been evaluated, for example, by the Defense Advanced Research Project Agency (DARPA) Urban Challenge and the Google self-driving car project. Results have demonstrated that the “rich map” approach to reliable autonomous navigation in structured, well-characterized terrain may be achievable; however, one of the substantial challenges to relying on a detailed prior map is that the system must maintain precise, accurate knowledge of the vehicle location within the map.
Traditionally, GPS receivers have been used to provide global localization; however, GPS alone does not provide the precision necessary to stay within a lane of traffic. Furthermore, GPS accuracy degrades severely in environments with multipath or shadowing and the signals can easily be blocked or intentionally disrupted by others desiring to interfere with operation. GPS-based systems can be improved with differential signal algorithms or by augmenting the location estimate with inertial sensors; however, the resulting systems are typically costly and may require additional infrastructure such as base stations.
To overcome these limitations, some autonomous vehicles sense the local environment and register the sensor measurements to a map of prior observations. The success of this map-matching approach depends significantly upon the type of sensor and the locale. Passive visual methods, for example, methods employing one or more digital video cameras, perform poorly in outdoor environments due to changes in scene illumination. Variations in solar illumination angle and cloudiness may result in a system seeing a substantially different scene than the scene sensed during prior map observations. Visual sensing is another approach and is used by the Google car. Rather than relying on natural or uncontrolled illumination, the sensor transmits light, typically at frequencies that are otherwise relatively dark, and measures the intensity of the return. Algorithms are used to search for similar intensity patterns in a map of previous measurements to determine the location of the vehicle. While the active sensing approach solves the problem of inconsistent scene illumination, difficulties are encountered when the apparent scene changes significantly with weather conditions. Snow, for example, obscures important features required to match a scene to prior acquired scenes. Other weather conditions such as fog, rain and dust also impact performance. Moreover, dynamic aspects of the observed scene can create further complications. For example, the presence of other moving vehicles and the motion of typically static features due to wind gusts present significant challenges. Thus, current sensing approaches for map-based localization are not robust to common real-world conditions. Furthermore, active sensors such as those employing light detection and ranging (LIDAR) sensors generally require expensive precision-engineered electro-optical-mechanical systems.